My current workflow: I describe the component in plain English with specific constraints ("a data table with sortable columns, sticky header, and virtual scroll

My current workflow: I describe the component in plain English with specific constraints ("a data table with sortable columns, sticky header, and virtual scrolling for 10k+ rows"), let the LLM generate the first pass, then manually fix the edge cases it always misses. The key insight I've found: LLMs are great at generating the 80% scaffolding but terrible at the 20% that makes UI actually feel good — animation timing, scroll behavior, focus management, accessibility edge cases. So I've stopped asking them for "production-ready" components and instead ask for "the boring structural parts" so I can focus on the interaction details that users actually notice.